Google DeepMind introduced Gemini 4 Carbon, a new flagship artificial intelligence model designed to find and fix software vulnerabilities, senior vice president Koray Kavukcuoglu said in a company blog post on October 28.
According to the post, the model detects vulnerabilities in real-world code with 88.2% accuracy on the SEC-bench Pro test and turns 54.8% of the flaws it finds into working attacks on ExploitGym. Its success rate at prompt injection, hiding commands in text to trick the model, fell to zero on the Gray Swan benchmark, down from 0.7% for Argon.
Google is rolling out Carbon in stages. Starting this week, it will go to a subset of cybersecurity specialists, then to partners in the Fairwind Program, and only later to developers and consumers. The company said the staged approach reflects its participation in the U.S. voluntary program for testing models before release.
Introductory pricing starts at $2 per million tokens sent to the model and $10 per million tokens it generates, with a 95% discount on cached tokens, the company said. Prices will rise to $4 and $20, respectively, once the introductory period ends.
On the Artificial Analysis Intelligence Index v5, Carbon scored 59 points, versus 50 for Argon, putting it in first place among all models. Its accuracy rose to 58% from 48%, but its hallucination rate climbed to 11% from 4%.
Google released Argon, the previous Gemini 4 flagship, on October 1 after a long delay.
